Unverified50% confidenceTradeoffExact time

Using a broader aggregation approach to improve coverage distorts benign prompts that are adjacent to safety.

1

Sources

50%

Confidence

Long-term

Relevance

10/5/2026

First Seen

Sources

CALM: A New Approach to Text-to-Image Safety — Local Counterfactual Corrections Break the Global Filtering Deadlock

Related Entities

safety-adjacent promptsconcept

Related Claims

Unverified聚合多个基准的评测方式降低了单一基准被针对性优化（刷榜）的风险67% similar UnverifiedSimply suppressing safety heads may risk a decline in safety, so dual-branch logits fusion is needed to balance usability and safety.67% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/975607

API

curl https://kongchang.com/api/v1/knowledge/claims/975607

MCP

get_claim(id=975607)